A satellite learned to find things on its own. The open question is the decision per watt.

In April 2026, a satellite did the thing.

NASA's Jet Propulsion Laboratory, Google DeepMind, and Loft Orbital flew Gemma 3 on an NVIDIA Jetson Orin aboard the YAM-9 satellite and let it answer natural-language queries in orbit — picking out infrastructure around railway hubs, classifying where natural land meets human development — with no ground analysts in the loop. JPL's software is called NAVI-Orbital. DeepMind's own framing of Gemma 3: "purpose-built for edge… designed to run on limited hardware far from a data center." The shift is from download-everything-then-triage to filter-then-transmit. (TechCrunch, 2026-06-15.)

That's a small local model, on a low-power board, deciding what matters where the cloud cannot reach. It is, almost exactly, the bet we've been measuring from a garage in Austin.

Find is the first half. Decide is the harder one.

What YAM-9 demonstrated is onboard perception — a vision-language model finding things in imagery. The layer immediately above it is onboard decision reasoning: given what was found, plus telemetry, plus hard physical constraints — what should the system do? A spacecraft on a 4-to-24-minute light delay, on a solar/RTG watt budget, running rad-hard compute generations behind your phone, has to reason about its own faults and choices by itself. Find → decide. The satellite flew the first half.

We'd already been benchmarking the second half — on the same class of hardware, with the number nobody else reports: decisions per watt.

On NVIDIA's own board, the most decisions per watt came from Google's model

A few days before this satellite story broke, we put three small models on a $250 Jetson Orin Nano Super and measured them on physics-grounded mission-decision scenarios — accuracy and tokens-per-second per watt:

• Gemma-4 E2B (DeepMind) — 2.0–2.2 tok/s/W, at every power mode, drawing the least power (~18 tok/s at 8.8 watts).
• Llama-3.2-3B (the model NVIDIA recommends for the board) — ~1.3.
• Nemotron-mini (NVIDIA's own) — ~1.1.

Gemma-4 E2B delivers 1.6–1.9× the tokens-per-watt of either NVIDIA-side model — the result that doesn't move with the prompt or the power mode. On decision accuracy, Gemma and Llama tie (~90–100%, the prompt decides which leads); NVIDIA's own model trails. The full benchmark, the raw per-item outputs, and the honest caveats (Ollama not TensorRT; n = 20; ground-analog, not flight-rated) are all open: github.com/SoftBacon-Software/low-power-edge-bench.

Why 8 watts is the whole point

Eight watts is a solar panel. A cubesat. A battery in a truck on the far side of a dead grid. When the data center is physically unreachable — orbit, deep space, a disaster zone — owned intelligence that runs on the watts you can actually generate stops being a slogan and becomes the only kind that works. Sovereignty, off-planet, isn't ideology. It's the design constraint. YAM-9 just made that concrete at 400 kilometers up.

The perception layer just flew. The decision-per-watt layer above it is wide open, the data's on the table, and we're measuring it in public. Run your own model. PR a contender. Tell us where we're wrong. That's still the bar.